WORKING PAPER Department of Economics Lennart Ziegler - What is the Media Impact of Research in Economics?

Page created by Daniel Lee
 
CONTINUE READING
Department of Economics
              WORKING PAPER

                 Lennart Ziegler
What is the Media Impact of Research in Economics?

             Working Paper No. 2103
                    July 2021
What is the Media Impact of Research in Economics?∗
                                           Lennart Ziegler∗∗
                                         University of Vienna

                                              July 2021

                                               Abstract

      Many research institutions aim to have a strong public impact but little evidence exists
      on the extent to which research findings reach a wider audience. Using a large sample of
      studies released in the working paper series of the National Bureau of Economic Research,
      I identify online coverage of research findings in 6 major news outlets. The analysis shows
      significant coverage rates in most newspapers in the first month after study release. Over-
      all, about every 11th working paper is covered at least once during this period. I also find
      that media reporting is correlated with several author and study characteristics. While
      differences in coverage between most research areas are modest, empirical as well as US-
      focused studies receive substantially more attention. In particular, widely cited papers
      are covered more frequently, showing that academic success of studies serves as a strong
      predictor for wider public impact.

      JEL Codes: A11, A14, L82
      Keywords: Economics research, media coverage, public impact

  ∗
     I am indebted to Dan Hamermesh who provided valuable comments and suggestions on an earlier draft of
this paper.
  ∗∗
     Address: University of Vienna - Department of Economics, Oskar-Morgenstern-Platz 1, 1090 Vienna,
Austria. E-mail: lennart.ziegler@univie.ac.at.

                                                   1
1    Introduction
Researchers often seek to communicate their research findings to a broader audience beyond
academic circles and contribute to the public debate. This is especially relevant at publicly
funded institutions, which should generate knowledge that benefits society as a whole. In
disciplines likes economics, research output can also inform policy makers regarding the im-
plementation of public policies and it can help voters to make more educated decisions. Yet,
little is known about the extent to which research findings reach a wider audience. While
citations and publication records help to determine the academic success of researchers, there
exist no comparable metrics to evaluate public outreach.
    One important channel to disseminate knowledge is media coverage of research output.
A discussion of new research findings in major news outlets allows to reach a much broader
audience than academic publications. Furthermore, research journalists act as an important in-
termediary between academics and society. They process complex analyses into non-technical
articles, which makes new findings also understandable for people without an academic back-
ground.
    Media exposure and academic success do not need to be strongly correlated. While jour-
nalist may rely on journal rankings and citations to judge the quality of a study (if already
available), interests and preferences may differ between academics and the broader public. Not
all papers with a high citation count will necessarily have a similar public impact.
    Despite the importance of media channels in spreading knowledge, little is known about
the extent to which economics research is discussed in the media and how exposure differs
between studies. This paper aims to fill this gap by examining the coverage of newly released
working papers in 6 major US news outlets. Using a large sample of studies released in the
working paper series of the National Bureau of Economic Research (NBER) between 2010 and
2019, I estimate the number of media citations and analyze how they differ by characteristics
of studies and authors.
    A key advantage of analyzing the coverage of working papers is that the immediate impact
can easily be measured. Before studies are eventually published in academic journals, they are
often circulated for the first time as working papers. Another important difference to journal
articles is that recently published working papers are neither cited nor published yet and thus
cannot affect the journalist’s decision to cover the paper.
    To obtain a measure of media exposure in the first weeks after the release of working papers,
I match author names to the online content of 6 widely read international news outlets (The
New York Times, The Washington Post, The Wall Street Journal, The Financial Times, The
Economist and CNN ). Estimation results show that the media outlets frequently cover research
in economics. Overall, an average working paper is matched to about 0.16 news articles in
the first 4 weeks after release. Every 11th study is covered in at least one of the news outlets

                                                2
during this period.
      Differences in coverage rates by JEL classification of research area are, in most cases,
relatively similar. Matching keywords in abstract texts, I find that empirical and US-focused
studies receive particular media attention. Even though associations with the journal impact
factor of eventually published papers are small, studies that are widely cited by the academic
community get more widespread media coverage. I interpret this as evidence that academic
success of a study serves as a strong predictor for wider public impact.
      Next to these study differences, media attention also varies with author characteristics.
Research strength of authors, as proxied by article views, is associated with higher coverage
rates. Yet, this effect is smaller and insignificant when I exploit within-study differences
of coverage between authors. Less experienced researchers and those affiliated to the top 5
percent institutions according to citations tend to get more coverage as well. These differences
remain significant when I account for study fixed effects. For a given working paper, media
articles are more likely to mention authors from research-strong institutions and those who
have less experience in academia.
      To my knowledge, this is the first study that quantifies the media exposure of economics
research and examines differences by study and author characteristics. Most closely related to
my analysis is an article by Hamermesh (2004), who provides media guidelines for economists
and includes anecdotal evidence on what sort of studies are typically covered by media outlets.
The paper highlights that interests of the public do not always align with academic interests.
Studies that reveal important facts to a broader audience might sometimes be too specific to
lay ground for further research. Furthermore, it is not clear how media attention affects the
chances to be published in a top-ranked journal.
      The methodological approach of this study bears similarity to a number of papers that
focus on the determinants of academic success in the economics profession. Examining pub-
lication trends over 6 decades, Hamermesh (2013) shows that journals increasingly publish
empirical studies with self-collected or experimental data. Card and DellaVigna (2020) ana-
lyze submissions to leading journals in economics, showing that referee and editor choices are
strong predictors of study citations. A related study by Card et al. (2020) provides further
evidence on gender discrimination in the publication process. The authors find that, con-
ditional on study citations, female authors have lower chances to get a paper published in
these journals.1 From a broader perspective, my analysis also relates to the large literature in
economics which examines how media channels influence politics and public policies.2
      The rest of the paper is organized as follows. Section 2 describes the different data sources
and provides descriptive statistics on working papers, authors and media coverage. The empir-
  1
     Bransch and Kvasnicka (2017) find no evidence that these gender differences are driven by publishing
boards consisting mainly of male editors.
   2
     See Prat and Strömberg (2013) for a comprehensive review of recent studies in political economy.

                                                   3
ical framework is outlined in Section 3. Section 4 presents estimates for overall media coverage
and shows how effects differ by study and author characteristics. Section 5 concludes.

2     Data
The estimation sample combines various data on media coverage, research papers and their
authors. It is composed from different sources of publicly accessible data. Table A.1 in the
appendix provides an overview with references to the respective sources and extraction dates.

2.1   Research output
The analysis focuses on all working papers which have been released by the National Bureau
of Economic Research (NBER) between 2010 and 2019. A key focus of this non-profit orga-
nization is to disseminate research among academics, policy makers and professionals. The
NBER working paper series is a prestigious and well-known source of recent research output.
About 20 papers are released every week and assigned to different research programs, which
cover all major fields in economics. Although access to NBER content is restricted, academic
institutions as well newspapers and magazines can view and download the working papers. As
such, the NBER publications are also frequently consulted by journalists.
    To publish an NBER working paper, at least one author has to be affiliated to the organi-
zation as faculty research fellow or research associate. Despite this restriction, many coauthors
are affiliated to non-US institutions and several studies focus on other countries than the US.
For this reason, the working paper series can be seen as a source of international research,
with impact also beyond the US.
    While the analysis could be replicated with other working paper series, focusing on NBER
working papers has several advantages. First, the organization is one of the largest and best-
known platforms for research in economics. With around 1,000 releases per year, the large
sample size simplifies the empirical analysis. Also, many journalists know the organization
and are thus aware of recent research output. Second, the NBER programs cover all major
areas of economics, which allows to examine differences in media coverage by subfield. Third,
the exact date of release is available. As explained in the next section, this greatly simplifies
matching papers to media reports and allows to estimate the immediate impact in the first
weeks after release.

2.2   Study and author characteristics
To study heterogeneous effects, I merge additional data from different online sources. The
NBER website provides for every working paper the date of release, JEL classification codes,
a publication record (if any reported) and the abstract text. The widely-used JEL codes,

                                               4
developed by the Journal of Economic Literature, provide a detailed classification scheme for
research articles in economics. Because of sample size restrictions, I focus on the 20 general
categories.3 While the JEL classification specifies the field of research in economics, the
codes do not distinguish many relevant key characteristics of studies. To obtain additional
information, I use word matches in the abstract text to proxy whether a study is empirical,
theoretical, policy related and US focused.4
       The NBER website also lists the researchers’ name and affiliation. From the profile pages
of authors, I obtain the year in which they released their first NBER publication. The years
between release of a working paper and first NBER publication will serve as proxy for research
experience in the subsequent analysis.
       To determine the gender of authors, I follow a procedure similar to the one proposed by
Card et al. (2020) and predict the gender based on first names. Using the R-package gender, I
obtain the proportion of females and males with a given name from US social security records
and from the Genderize.io database, which is compiled from user profiles on social networks.
Because random measurement error in the explaining variable will downward bias regression
coefficients, it is necessary to avoid mix-ups as much as possible. Therefore, I only classify
the gender if the proportion of females or males is at least 90 percent. Overall, 92 percent
of all names in the estimation sample can be identified as male or female. As a robustness
check of the described procedure, I search for author matches in the RePEc ranking of the
top 10 percent female economists and on a list of female members of the European Economic
Association (EEA). Among the matched authors, only 2 percent of female economists are
wrongly classified as males.
       Additional information on research output and authors comes from the RePEc (Research
Papers in Economics) platform. The RePEc service EconPapers provides a large collection of
working papers and journal articles in economics. For each working paper, I obtain the total
number of citations.5 In addition, I count the number of additional working and discussion
paper series in which the studies were released. Multiple releases can increase the visibility of
a paper and potentially lead to more media coverage.
       To evaluate the research strength of journals, institutions and authors, I collect data from
the RePEc services IDEAS and CitEc, which provide various citation-based rankings. The
CitEc journal ranking lists more than 600 journals along with their impact factor, which can
be used as a measure of publication success. The IDEAS institution ranking contains the top
5 percent of institutions according to overall citations. Furthermore, I extract the number of
article views for all authors who are registered on RePEc.
   3
     Note that these groups are non-exclusive. In fact, most papers are classified as belonging to multiple
categories.
   4
     See Figure 4 for the definition of word pattern matches.
   5
     If a working paper has already been published or if other working papers exist, I consider the total number
of citations across all publications.

                                                       5
Table 1: Descriptive statistics - Working papers

                                                             Mean     Std. Dev.    Min     Max        Obs
   # authors                                                  2.60       1.09        1      13       10,902
   Single author                                              0.14       0.34        0      1        10,902
   Publication
   Citations                                                 25.93      56.36        0     1,493     10,902
   Published (yet)                                            0.63       0.48       0         1      10,902
   Top five journal                                           0.18       0.39       0        1        6,903
   Journal impact factor                                      4.18       3.24      0.02    12.59     5,522
   # other WP versions                                        0.80       1.23       0        11      10,902
   JEL codes
   (J) Labor and Demographic Economics                        0.24       0.43        0       1       10,902
   (E) Macro-/Monetary Economics                              0.23       0.42        0       1       10,902
   (I) Health, Education, and Welfare                         0.22       0.42        0       1       10,902
   (D) Microeconomics                                         0.22       0.42        0       1       10,902
   (G) Financial Economics                                    0.21       0.40        0       1       10,902
   (H) Public Economics                                       0.17       0.38        0       1       10,902
   (O) Econ. Develop., Techn. Change, and Growth              0.17       0.37        0       1       10,902
   (F) International Economics                                0.15       0.36        0       1       10,902
   (C) Mathematical and Quantitative Methods                  0.10       0.30        0       1       10,902
   (L) Industrial Organization                                0.11       0.31        0       1       10,902
   Other (
As shown in Table 2, the estimation sample contains studies written by about 9,600 dif-
ferent authors. The majority publishes only one working paper during the sample period but
there also exist some authors who publish very frequently in this series, yielding an average of
almost 3 papers per author. Among the 92 percent of names that can be matched to a gender,
a quarter of authors are female.6 56 percent of authors have a matching RePEc profile, which
reports overall article views. Moreover, 59 percent of authors are affiliated to an institution
that is among the top five percent in the IDEAS citation ranking.

                                Table 2: Descriptive statistics - Authors

                                                Mean         Std. Dev.        Min       Max         Obs
         # WP (sample period)                    2.94           4.32           1        75         9,634
         Year first NBER WP/Pub               2,011.04          8.07         1,973     2,019       9,634
         Female                                  0.26           0.44           0         1         8,840
         Gender unknown                          0.08           0.28           0         1         9,634
         Author RePEc profile                    0.56           0.50           0         1         9,634
         Author RePEc article views           25,270.61      47,337.82         0      674,785      5,396
         Insitution in top 5% cited              0.59           0.49           0         1         9,634
         Institution citation score          112,548.26      119,804.41      6,645    470,596      5,729
         Note: The sample includes authors of NBER working papers released between 2010 and
         2019. Data on article views and citation scores were obtained in May/June 2021.

2.3      Media coverage
To measure media coverage, I collect data from the websites of four daily newspapers (The
New York Times, The Washington Post, The Wall Street Journal, The Financial Times),
one weekly magazine (The Economist) and one television channel (CNN).7 I focus on these
6 sources because they all have substantial reach even beyond the US and report extensively
on the economy and related topics. Being discussed in one of the outlets guarantees research
papers to have a widespread audience. Even though this only represents a small share of
potential media coverage, these sources thus serve as a reasonable proxy for media exposure.8
The observed media content (reports, features, etc.) often overlaps with print content but
   6
      Card et al. (2020) compute a share of about 18 percent female authors of publications in 53 different
journals during a similar time period.
    7
      The NBER website also lists news articles that cover the working papers. While these self-reported entries
include a wider range of news outlets, they only cover a selective subset of media coverage. The number of entries
substantially increases during the period of observation, which must be due to selective reporting. Moreover,
it is possible that the selection of reported coverage is correlated with author and paper characteristics. Also,
the entries do not readily allow to identify which authors of a study are mentioned in the media.
    8
      The sample of newspapers can be further extended. Yet, not all websites qualify for web scraping in this
context. The respective website has to have a reliable search engine which allows for whole word search and an
accessible archive of published articles in the past years along with publication dates. These criteria exclude
a number of potential news sources (e.g. Los Angeles Times, Chicago Tribune, FoxNews, CBS News, USA
Today or Time Magazine).

                                                        7
also includes additional online-only content such as affiliated blog posts. To match media
reports to papers, I use the websites’ search engines to search for the names of all authors
in the sample. Articles or blog entries that discuss a research paper typically mention the
corresponding authors, which can be used to identify studies.9 This simple procedure also
allows to automatize the search process. Due to the large number of working papers and
authors, a manual search for media coverage would hardly be feasible. A caveat of this
approach is that I cannot distinguish the extent or type of coverage. While some articles
provide an in-depth discussion of research findings, others may only make a short reference to
working papers.
      The main piece of information contained in the search results are the publishing dates
of news articles. Not all observed name matches relate to news coverage of the respective
studies. To identify coverage of working papers, I exploit the exact timing of release and
compare matches in the first weeks after release to matches in the weeks before. Figure 1 plots
the number of weekly matches per working paper by study release date, starting 6 months
before and up to 6 months after release. The upper solid line shows the number of matches for
the full sample. A news article is counted as match if it refers to at least one of the authors.
In all weeks, I observe a large number of matches to working papers in my sample. Following
the release of the paper, media hits increase sharply by more than 80 percent. After a few
weeks, matches drop again to numbers similar to pre-release levels. While the large spike can
arguably be attributed to coverage of the respective working papers, most of the matches,
especially before the release dates, should be unrelated news articles.
      When I only count matches if at least two authors of a study are referenced in a news
article, the number of potential mismatches can be substantially reduced. By construction,
this excludes single-authored working papers, which amount to 14 percent of the sample. The
lower graph of Figure 1 shows that the corresponding number of matches is close to zero in
most of the weeks. Almost all coverage happens in the first weeks after release. Compared
to the unrestricted sample, the peak is somewhat less pronounced. This suggests that not
all news articles provide full references when reporting on recently released working papers.
To capture the full extent of media coverage, the subsequent analysis will focus on the full
sample. Corresponding results for the restricted sample, which tend to be estimated with
smaller standard errors due to a lower number of mismatches, are reported for comparison in
the appendix.
      For matches that are unrelated to coverage of a given working paper, a higher number of
authors should increase the chances to observe a match. The dashed line plot of Figure 1 shows
matches to multi-authored studies if news articles are also counted when only one author is
  9
    For a small number of authors, I observe many mismatches because they share their name with other
persons who are frequently mentioned in news sources. To reduce measurement error, I drop all working
papers written by these authors, which corresponds to less than 1 percent of the sample.

                                                 8
mentioned. Compared to the full sample of papers, there are indeed somewhat more media
matches. To account for this effect, I will control for the number of authors in the subsequent
analysis.

                                            Figure 1: Matches by week of release (by sample)
                                                                                     Mult. authors & mult. mention
                                     0.15
                                                                                     Mult. authors & any mention
                                                                                     All authors & any mention
       # matches per working paper

                                     0.10

                                     0.05

                                     0.00
                                              −20         −10             0            10             20
                                                                Week after release

3     Empirical framework
3.1   Measurement error
Using automated search of author names to identify news coverage poses several empirical
challenges. As shown above, the procedure may wrongly attribute some news articles to
research papers. In some cases, the mentioning of an author even coincides with the release
of a working paper although the media coverage is unrelated to the respective study. This
especially concerns economists who regularly write op-ed’s or blogs. It is also possible that
other unrelated persons with the same name are referenced instead. Moreover, media sources
may report about a study without mentioning the full or correct name of authors. Both types
of measurement error can bias the estimated treatment effect. In the following, I assume that,
                                                ∗ is measured with error  :
for study i in period t, the number of matches yit                        it

                                                                     ∗
                                                              yit = yit + it

To identify coverage of working papers, I exploit the timing of release and compare matches
in the first weeks after release to matches in the weeks before. The actual treatment effect on

                                                                    9
the number of matches is

                                          ∗                 ∗
                                   β = E[yit |Dit = 1] − E[yit |Dit = 0],

where Dit indicates post-release periods. The estimated treatment effect is

            E[β̂] = E[yit |Dit = 1] − E[yit |Dit = 0] = β + E[it |Dit = 1] − E[it |Dit = 0].

The identifying assumption thus requires that the average error is the same before and after
the release of a working paper:

                                      E[it |Dit = 1] = E[it |Dit = 0]

This condition is likely satisfied when the mismeasurement can be attributed to over detection.
Unless the author publishes multiple studies at the same time or suddenly becomes more
present in the media, the extent of measurement error will not be affected. Also matches to
other persons with the same name should not be correlated with the release date. However, if
a media outlet covers a new working paper without mentioning the authors, this only leads to
measurement error in post-release periods. As a result, the estimated treatment effect will be
downward biased. Because it is common journalistic standard to reference at least one author
when research studies are discussed in the media, I expect the resulting bias to be small.
      Next to the number of matches, it is also informative to measure whether a study is
covered at all in the media. Identifying this effect requires additional assumptions. The actual
treatment effect on having any match is given by

                                      ∗                     ∗
                              γ = P [yit > 0|Dit = 1] − P [yit > 0|Dit = 0],

while the observed difference in the share of non-zero matches before and after release is

                                   ∗                           ∗
                       E[γ̂] = P [yit + it > 0|Dit = 1] − P [yit + it > 0|Dit = 0].

                           ∗ and  are independent, this can be expressed as
Assuming that it ≥ 0 and yit     it

                                   E[γ̂] = γ × (1 − P [it > 0|Di = 1]).

                                 ∗ and  , errors caused by mix-ups with other persons of the
To satisfy independence between yit     it
same name should not be correlated with coverage of the working paper. Furthermore, authors
of covered studies should not be more present in the media either. Under these conditions,
the measurement error biases the actual impact towards zero as shown above.10 Assuming
 10
      The derivation is provided in the appendix.

                                                     10
that an error is equally likely before and after study release, the matching probability in the
pre-release period can be used to correct the downward bias.

3.2   Event study
To estimate media exposure of research papers by week after release, I conduct an event study.
The estimation equation is given by

                                                12
                                                X
                                                           w
                                  yit = αt +           βw Dit + uit
                                               w=−12

                                                                          w indicate the week
where yit denotes the number of matches for study i in week t. Variables Dit
relative to the date of release, where the omitted indicator is the week just before release. To
account for potential timing effects, I also include release date indicators (αt ). This regression
allows to quantify news coverage by week before and after the release of a research paper.
Coefficients βw capture the increase in media matches relative to the base period. Under the
conditions specified above, this should remove the measurement error. Furthermore, the event
study design allows to reveal potential trends in the pre-release period.

4     Results
4.1   Overall coverage
Figure 2 plots estimates of the event study, along with 95 percent confidence intervals, for
the full sample of working papers. The graph shows a large spike in coverage immediately
after release. In the first week, I estimate about 0.07 matches per study. Afterwards, coverage
rates drop again but remain significant for another two weeks. In the subsequent weeks,
coverage further decreases to estimates slightly above zero. Compared to the base period,
coefficients are also close to zero in the weeks before the release. This suggests that researchers
do not disseminate their research findings prior to the NBER release, which allows a clear
identification of coverage around the release date.
    To simplify the analysis in the remainder of this paper, I measure media coverage in the
first 4 weeks after the release of working papers (treatment period) and compare it to matches
in the preceding 4 weeks (control period). Table 3 reports the corresponding estimates for
total coverage and coverage by media source. As shown in the upper panel, I observe nearly
0.16 media hits per paper. A lot of coverage stems from The Wall Street Journal, which
accounts for half of the matches. I also find significant estimates for most of the remaining
media outlets, ranging from about 0.003 to 0.024 matches per paper.

                                                 11
Figure 2: Event study graph

                                            0.08

                                                                                                        ●
         # matches (relative to week −1)

                                            0.06
                                                                                                            ●

                                            0.04

                                            0.02
                                                                                                                ●

                                                                                                                                                ●
                                                                                                                        ●                   ●
                                                                                                                            ●                           ●
                                                                                                                    ●
                                                                                  ●                                                     ●
                                                                 ●                                                              ●
                                            0.00    ●
                                                             ●                             ●
                                                                                                    ●                                               ●
                                                                     ●    ●           ●                                             ●
                                                                              ●
                                                                                               ●
                                                         ●

                                           −0.02

                                                   −12               −8               −4                0               4               8               12

                                                                                          Week after release
                                                                                  Note: The grey area indicates 95% confidence intervals.

   The lower panel of Table 3 reports estimates for the likelihood to observe any match. As
shown in the previous section, calculating the difference between treatment period and control
period in the share of papers that have any match yields a downward biased estimate for the
coverage likelihood. To correct for measurement error, I divide the difference by the share of
                                                               D=1 −Share(y>0)D=0
papers that are not matched in the control period ( Share(y>0)
                                                          1−Share(y>0)D=0         ). The results
show that approximately 9 percent of studies were covered in at least one of the 6 media
sources. Again, I find the highest probability of coverage for The Wall Street Journal.
   To examine changes in coverage over time, I compute average differences in media matches
between treatment period and control period by month of release. Figure A.1 in the appendix
shows that there is no clear trend in coverage over the sample period. I do not find evidence
for seasonality in news coverage either. The estimates do not differ significantly by months,
weeks or weekdays.

                                                                                               12
Table 3: Coverage estimates by sample

                            All       NYT         WP       Economist        WSJ         FT        CNN
      Outcome: Number of matches
      Coefficient     0.158   0.006   0.022                   0.024        0.085      0.018       0.003
      Standard error (0.012) (0.005) (0.005)                 (0.003)      (0.006)    (0.004)     (0.001)
      Outcome: Any match
      Coefficient     0.087   0.008              0.020        0.019        0.058      0.009       0.003
      Standard error (0.005) (0.003)            (0.003)      (0.002)      (0.003)    (0.002)     (0.001)
      Note: The sample includes all papers (10,902). The upper panel reports differences in media
      matches between the first 4 weeks after the NBER release date (release period) and the 4 weeks
      before (control period). Standard errors reported in parentheses are clustered by paper. The lower
      panel shows differences in the share of papers with any match between release and control period
      divided by the share of papers without matches in the control period, where standard errors are
      approximated using the delta method.

   Coverage rates estimated on the restricted sample, where news articles are only counted
if they can be matched to at least two authors, are reported in the appendix (see Table A.2).
For better comparison, the table also provides estimates for the sample of multi-authored
papers, where news articles are also counted when only one author is mentioned. Compared
to the results shown above, the estimate for the overall number of related news articles is
about 30 percent lower in the restricted sample. Systematic measurement cannot explain this
difference because the correlation between release date and measurement error would have to
be unrealistically large. It is unlikely that authors are suddenly much more present in the
media after the release of one study. The observed difference rather indicates that some news
articles do not provide complete references. As a result, I observe a smaller number of matches
in the restricted sample. Because of reduced measurement error, the corresponding t-statistics
tend to be smaller as well.

4.2   Differences by study characteristics
Not all working papers receive the same degree of media attention. This section analyzes to
what extent coverage differs by a number of observable study characteristics such as academic
popularity and field of research. As discussed in the previous section, the estimated standard
errors in this analysis are relatively large because many media matches in the full sample
are unrelated to coverage of the working papers and have to be differenced out. The more
restrictive definition, which requires that media articles mention at least two authors, misses
out some related matches but allows for a more precise estimation of heterogenous effects.
Tables and figures on heterogeneous effects for the full sample are presented in the main
text while differences by study characteristics for the restricted sample are provided in the
appendix.

                                                      13
Figure 3: Media reports by JEL classification

                 (J) Labor and Demographic Economics

                       (E) Macro−/Monetary Economics

                                   (D) Microeconomics

                                          Other (
Figure 4: Media reports by study characteristics

                                    Text pattern matches
                              0.3   Empirical: 'data', 'empiric*', 'estimat*', 'experiment*'
                                    Theoretical: 'theor*', 'model'
                                    Policy: 'polic*'
                                    USA: 'United States', 'US*', 'America*', US states & 100 largest cities
            # media matches

                              0.2

                              0.1

                              0.0

                                          All         Theoretical       Empirical           Policy            USA
                                                       (42.32%)         (62.11%)          (23.43%)          (19.86%)

                                                                     Note: The black lines indicate 95% confidence intervals.

reference, the estimated coverage rate is more than 60 percent higher. Although all 6 media
outlets are internationally known, many readers are based in the United States. Journalists
might thus be more likely to discuss a study when there exists a connection to the United
States.
    Next, I examine the relation between media coverage and academic success. Table 4 reports
differences in coverage by citations of a study and, if the paper has already been published
in a ranked journal, by journal impact factor and journal rank. To account for the fact
that recently released papers are less likely to be published yet, all regressions include release
date indicators. None of the estimated specifications reveal significant differences in coverage
between published and unpublished studies.
    As shown in the first column of Table 4, there exists a robust correlation between media
reports and academic citations. A doubling of citations is associated with 0.055×ln(2) ≈ 0.038
additional media reports per study in the first weeks after release.11
  11
     To also include papers without any references, I add one unit to every citation count before taking the
logarithm. Card et al. (2020) use an inverse-hyperbolic-sine transformation, asinh(cite) ≡ log(cite + (1 +
cite2 )0.5 ), to accommodate observations with zero citations. Applying this transformation leads to similar
coefficient estimates.

                                                                       15
Table 4: Media reports by study characteristics

                                                     (1)         (2)         (3)         (4)          (5)
      WP released

      × ln(Citations+1)                          0.055***                            0.061***      0.060***
                                                  (0.011)                             (0.012)       (0.012)
      × ln(Journal impact factor)                               0.015                  -0.012
                                                               (0.026)                (0.028)
      × ln(Journal rank)                                                   -0.015                    0.003
                                                                          (0.012)                   (0.013)
      × I(Ranked publication observed)                          0.077      0.053       -0.011        0.040
                                                               (0.133)    (0.048)     (0.139)       (0.048)
      Note: The sample includes all papers (10,902). The coefficients measure media matches in the
      first 4 weeks after the NBER release date (release period) compared to the 4 weeks before (control
      period). All regressions control for levels of included characteristics, date of release, number of
      authors and JEL categories. Journal impact factors and ranks are set to the minimum if no (ranked)
      publication exists. Standard errors reported in parentheses are clustered by paper. I( ) denotes
      indicator functions. ∗ significant at 10% level, ∗∗ significant at 5% level, ∗ ∗ ∗ significant at 1%
      level.

   When using the logarithm of the journal’s impact factor or of the journal rank as proxy
for academic success, estimated differences are small and statistically insignificant. For the
restricted sample, I find very similar coefficients, which are significant at the 10 percent level
(Table A.3). For a 100 percent increase in the journal’s impact factor, measured media reports
only increase by about 0.01. This shows that citations are a much stronger predictor for media
coverage than journal strength.
   Both measures of academic success, citations and journal rank, are highly correlated
(Corr[ln(Journal rank), ln(Citations + 1)] ≈ −0.37). Since citations are also observed af-
ter publication, appearing in a high-ranked journal might drive up the citation count. When
I include both study citations and journal rank or journal impact factor in the regression, the
coefficient on citations remains similar while differences by journal strength get even smaller.
Results for the restricted sample show that the significant associations with journal rank and
journal impact factor fade to zero once citations are included in the regression.
   Given the lack of exogenous variation in coverage, it is unclear what drives the correlation
between media exposure and academic success. Being discussed in the media might raise
the chances of publication in a high-ranked journal. It is possible that editors prefer to
publish papers with media coverage as public attention increases the visibility of the journal.
Analogously, some researchers might cite a specific study because they have read a news
article about it. It is also likely that academic and media interests align to a certain extent.
Journalists and academics might have the same unobserved ‘taste’ for specific studies that are

                                                        16
innovative or comprise novel findings. The estimation results discussed above suggest that the
latter channel is more important. When controlling for citations, differences by journal rank
fade to be significant. If media coverage increases the publication success of papers conditional
on their academic value as proxied by citations, this association should persist.
   Dissemination of research in multiple working paper series could substantially raise the
visibility of a study in the media. To test this hypothesis, I estimate differences in coverage
rates by the number of working paper versions listed on EconPapers. The regression results of
Table 5 show that working papers released in multiple series tend to receive somewhat more
media exposure but the effect is small and insignificant. One additional release is associated
with an increase of about 0.01 media reports. The positive estimate is driven by studies with
multiple other releases, which only concerns a small share of working papers in the sample.
Less than 20 percent of studies have more than one additional release in other working paper
series. The corresponding estimates for the restricted sample (Table A.4) are again more
precisely estimated and confirm the findings for the full sample.

                Table 5: Media reports by number of working paper versions

                                                  (1)         (2)        (3)           (4)
               WP released

               × # other WP versions             0.012      0.008
                                                (0.012)    (0.012)
               × 1 other WP version                                     -0.015       -0.022
                                                                       (0.030)      (0.031)
               × 2 other WP versions                                    0.057        0.049
                                                                       (0.041)      (0.042)
               × 3+ other WP versions                                   0.024         0.011
                                                                       (0.044)      (0.046)
               JEL Categories                                 X                         X
               Note: The sample includes all (10,902). The coefficients measure media
               matches in the first 4 weeks after the NBER release date (release period)
               compared to the 4 weeks before (control period). All regressions control for
               levels of included characteristics, date of release and number of authors. Stan-
               dard errors reported in parentheses are clustered by paper. ∗ significant at
               10% level, ∗∗ significant at 5% level, ∗ ∗ ∗ significant at 1% level.

4.3   Differences by author characteristics
Next to study characteristics, differences between authors may influence media exposure as
well. Some researchers might be more likely to seek media attention than others. Visibility of
an author may also be an important criterion for journalists to decide which study is worth

                                                     17
covering.
       Table 6 shows how media coverage differs by gender, experience, research strength and
affiliation of authors. In contrast to the preceding analysis of study-level differences in coverage,
media matches are now measured for every author-paper combination. This sample allows
to examine whether studies are more likely to be covered if their authors possess specific
characteristics. The analysis further shows which authors are mentioned when news articles
do not reference all authors of a study. To capture heterogeneity with respect to the mentioning
of specific authors, I estimate regressions that also include study indicators.
       The results show no significant differences in media coverage between male and female
authors. Compared to studies written by male economists, the difference is at most 0.01
and never statistically different from zero. Likewise, the results do not provide evidence that
gender is a relevant factor when media reports only cite specific authors of a study. The point
estimates remain similar when study fixed effects are included in the regression.
       As measure of research strength, I use the authors’ total number of article views on Econ-
Papers, which yields a significant and robust coefficient. One standard-deviation increase in
ln(article views) is associated with approximately 0.03 additional media matches in the re-
gressions without study fixed-effects. Relative to the average number of media reports, this
corresponds to an increase of about 20 percent. When study indicators are added to the re-
gressions, the coefficients are only half as large and turn insignificant. This shows that the
differential impact by articles views of authors can mostly be attributed to correlated study
differences. When reporting on a specific working paper, media outlets are not significantly
more likely to mention those researchers who are more visible in academia.12
       Even if authors are individually less visible, journalists might be more likely to discuss their
research if they are affiliated to institutions with a strong research background. The regression
estimates show that authors affiliated to the top 5 percent institutions in the citation ranking
receive significantly more coverage. This effect cannot be explained by study differences,
suggesting that journalists rather cite those authors who are affiliated to a ranked university.
Conditional on being in the ranking, the institutions’ total citations are not associated with
more coverage. The estimated coefficients are small and insignificant in both specifications.
       Moreover, I estimate a negative correlation between media coverage and research experience
of authors. For every ten years between study release and an author’s first NBER publication,
coverage decreases by about 0.035 reports per release when I control for author visibility. In
the regressions with study fixed effects, the estimate decreases again but remains significant,
showing that less experienced authors are more likely to be referenced in news articles about
a working paper release.
  12
    Authors without a RePEc profile, for whom article views are not available, show higher coverage rates.
Yet, it is unclear how this effect should be interpreted as researchers who lack such a profile do not necessarily
have less known research output.

                                                       18
As for the previous section, I repeat the heterogeneity analysis for the restricted sample
of multi-authored papers, where news articles are only counted if they mention at least two
authors. Because there exists little within-paper variation in media coverage of authors in
this sample, the estimated specifications do not include study fixed effects. Table A.5 in
the appendix reports the corresponding estimates. Compared to the results shown in Table
6, the coefficients are mostly smaller in magnitude but qualitatively similar. Due to fewer
mismatches in this sample, the point estimates are again more precisely estimated. In contrast
to the coefficients for the full sample, differences by the citation score of ranked institutions
are marginally significant.

                               Table 6: Media reports by author characteristics

                                                  (1)           (2)            (3)            (4)            (5)            (6)
 WP released

 × Gender - Female                               -0.009       -0.008          -0.005        -0.007         -0.006        -0.008
                                                (0.012)      (0.013)         (0.012)       (0.013)        (0.012)       (0.013)
 × Gender - Unknown                              -0.009       -0.017          -0.005        -0.016         -0.005        -0.017
                                                (0.016)      (0.015)         (0.016)       (0.015)        (0.016)       (0.015)
 × Experience (10−1 ×years)                     -0.012*     -0.017***      -0.034***      -0.025***      -0.035***     -0.027***
                                                (0.006)      (0.006)         (0.008)       (0.008)        (0.008)       (0.008)
 × ln(Author article views)                                                 0.021***         0.008       0.019***        0.006
                                                                             (0.006)       (0.006)        (0.006)       (0.006)
 × I(Author article views observed)                                        -0.188***        -0.071       -0.170***       -0.062
                                                                             (0.053)       (0.055)        (0.051)       (0.054)
 × ln(School citation score)                                                                               0.005         -0.001
                                                                                                          (0.006)       (0.006)
 × I(School citation score observed)                                                                      0.032**       0.037**
                                                                                                          (0.015)       (0.017)
 Study FE                                                        X                            X                             X
 Note: The sample includes all author-paper combinations of multi-authored papers (28,326). The coefficients measure
 media matches in the first 4 weeks after the NBER release date (release period) compared to the 4 weeks before (control
 period). All regressions control for levels of included characteristics, date of release, number of authors and JEL categories.
 Experience is proxied by years since first NBER publication. Article views and citation scores are set to the respective
 minimum if author/school data not available. I( ) denotes indicator functions. ∗ significant at 10% level, ∗∗ significant
 at 5% level, ∗ ∗ ∗ significant at 1% level.

5     Conclusion
A key challenge for researchers is to make their knowledge accessible to a broader audience
outside academia. This paper provides evidence that top media outlets serve as an important

                                                                19
intermediary between academia and the general public. Analyzing a large sample of working
papers in economics released between 2010 and 2019, I find that new research findings are
widely covered in leading newspapers. About 9 percent of studies are mentioned in at least
one of 6 major news outlets in the first 4 weeks after release of a working paper.
   While I observe media coverage, it is unclear how media outlets become aware of interesting
studies. Journalists covering economics and related fields may consult publications of working
paper series and decide what is worth reporting on. Sometimes researchers also directly seek
contact to journalists to publicize new findings. Given the lack of empirical evidence on
media coverage of research, it is difficult to judge which studies are of particular interest for
the media. Although my estimates cannot be interpreted as causal differences, the findings
show that several study and author characteristics serve as important predictors of media
coverage. In particular, I find a strong correlation between media coverage and academic
success as measured by citations. This indicates that many studies which are popular among
academics are also of interest for the wider public. Moreover, I observe that media attention
varies with author characteristics. When reporting on a working paper, journalists are more
likely to mention those authors who have less experience and are affiliated to research-strong
universities. Researchers with these characteristics might seek more media attention or could
more likely be the corresponding author. Also, journalists might have a preference to cover or
interview researchers with such a background.
   A key question remains whether news coverage itself positively affects academic impact.
As most of the media articles appear right after the initial release of a working paper, it is not
possible to exploit differential timing of coverage. Future research may look at ways to exploit
credible exogenous variation in media exposure.

                                               20
References
Bransch, F. and Kvasnicka, M. (2017). Male gatekeepers gender bias in the publishing process?
  IZA Discussion Paper.

Card, D. and DellaVigna, S. (2020). What do editors maximize? Evidence from four economics
  journals. Review of Economics and Statistics, 102(1):195–217.

Card, D., DellaVigna, S., Funk, P., and Iriberri, N. (2020). Are referees and editors in eco-
  nomics gender neutral? Quarterly Journal of Economics, 135(1):269–327.

Hamermesh, D. S. (2004). Maximizing the substance in the soundbite: A media guide for
  economists. Journal of Economic Education, 35(4):370–382.

Hamermesh, D. S. (2013). Six decades of top economics publishing: Who and how? Journal
  of Economic Literature, 51(1):162–72.

Prat, A. and Strömberg, D. (2013). The political economy of mass media. Advances in
  economics and econometrics, 2:135.

                                             21
Appendix
Measurement error in the impact on having any match
Assumptions:

  1. The error is non-negative (it ≥ 0) and unrelated to the extent of coverage
       ∗ and  are independent).
     (yit     it

  2. There exists no coverage before the release (P [yi∗ > 0|Di = 0] = 0).

  3. The error probability does not change with the release
     (P [it = 0|Dit = 0] = P [it = 0|Dit = 1]).

Derivation:
            ∗                           ∗
E[γ̂] = P [yit + it > 0|Dit = 1] − P [yit + it > 0|Dit = 0]
     (1)        ∗                                             ∗
      = 1 − P [yit = 0|Dit = 1]P [it = 0|Dit = 1] − (1 − P [yit = 0|Dit = 0]P [it = 0|Dit = 0])
     (2)    ∗                                         ∗
      = P [yit = 0|Dit = 0] P [it = 0|Dit = 0] − P [yit = 0|Dit = 1]P [it = 0|Dit = 1]
        |        {z       }
                      =1
                                      ∗
     = P [it = 0|Dit = 0] − (1 − P [yit > 0|Dit = 1])P [it = 0|Dit = 1]
     (3)                                               ∗
      = P [it = 0|Dit = 0] − P [it = 0|Dit = 1] +P [yit > 0|Dit = 1]P [it = 0|Dit = 1]
        |                   {z                  }
                                =0
     (3)       ∗
      =    P [yit   > 0|Dit = 1]P [it = 0|Dit = 0]
     = γ × P [it = 0|Dit = 0]

                                                      22
Table A.1: Overview data sources

Type                      Source               URL                                          Extraction date

NBER:
Author profile            NBER website         www.nber.org
                                                                                               May 2021
Paper characteristics     NBER meta data       www2.nber.org/wp_metadata

Media:
New York Times            NYT website          www.nytimes.com
Washington Post           WP website           www.washingtonpost.com
Economist                 Economist website    www.economist.com                               May 2021

Wall Street Journal       WSJ website          www.wsj.com
Financial Times           FT website           www.ft.com
CNN                       CNN website          www.cnn.com

Additional information:
Journal ranking 2020      CitEc website        https://citec.repec.org
Institution ranking       IDEAS website        https://ideas.repec.org
Author article views      LogEc website        https://logec.repec.org
                                                                                            May/June 2021
Working paper versions
                          EconPapers website   https://econpapers.repec.org
Study citations
R-package ’gender’        CRAN                 https://cran.r-project.org
Top female economists     IDEAS website        https://ideas.repec.org/top/top.women.html
Female EEA members        EEA website          http://www.eeassoc.org

                             Table A.2: Coverage estimates by sample

                        All    NYT       WP   Economist   WSJ                        FT       CNN
       All matches - All papers (N =10,902)
       Coefficient     0.158   0.006    0.022   0.024     0.085                    0.018      0.003
       Standard error (0.012) (0.005) (0.005)  (0.003)   (0.006)                  (0.004)    (0.001)
       All matches - Multiauthor papers (N =9,429)
       Coefficient     0.162   0.009    0.024   0.025     0.086                    0.015      0.003
       Standard error (0.012) (0.005) (0.005)  (0.003)   (0.006)                  (0.004)    (0.001)
       At least two matches - Multiauthor papers (N =9,429)
       Coefficient     0.107   0.007    0.017   0.014     0.062                    0.007      0.000
       Standard error (0.006) (0.002) (0.002)  (0.002)   (0.004)                  (0.001)    (0.000)
       Note: The coefficients measure media matches in the first 4 weeks after the NBER release date
       (release period) compared to the 4 weeks before (control period). Standard errors reported in
       parentheses are clustered by paper.

                                                     23
Figure A.1: Name matches by month of working paper release

                                                                75

    Difference # matches (4 weeks post − 4 weeks pre release)

                                                                50

                                                                25

                                                                0

                                                                 2010     2011       2012      2013      2014      2015    2016     2017    2018    2019
                                                                                                             Month of release
                                                                                                        Note: The bold line shows smoothed values with 95%
                                                                                                         confidence intervals obtained from local regressions.

Figure A.2: Media reports by JEL classification (Restricted sample)
                (J) Labor and Demographic Economics

                                                                 (E) Macro−/Monetary Economics

                                                                              (D) Microeconomics

 (O) Econ. Develop., Techn. Change, Growth

                                                                      (F) International Economics

                                                                         (G) Financial Economics

                                                                                     Other (
Figure A.3: Media reports by study characteristics (Restricted sample)
                              Text pattern matches
                              Empirical: 'data', 'empiric*', 'estimat*', 'experiment*'
                              Theoretical: 'theor*', 'model'
                       0.15   Policy: 'polic*'
                              USA: 'United States', 'US*', 'America*', US states & 100 largest cities
     # media matches

                       0.10

                       0.05

                       0.00

                                    All         Theoretical         Empirical         Policy            USA
                                                 (42.59%)           (64.08%)         (22.3%)          (19.59%)

                                                               Note: The black lines indicate 95% confidence intervals.

                 Table A.3: Media reports by study characteristics (Restricted sample)

                                                              (1)          (2)         (3)           (4)            (5)
WP released

× ln(Citations+1)                                        0.041***                                0.042***       0.042***
                                                          (0.006)                                 (0.006)        (0.006)
× ln(Journal impact factor)                                              0.017*                    -0.001
                                                                         (0.010)                  (0.010)
× ln(Journal rank)                                                                  -0.012*                       0.000
                                                                                    (0.006)                      (0.006)
× I(Ranked publication observed)                                          0.067      0.019         0.003          0.009
                                                                         (0.048)    (0.020)       (0.048)        (0.020)
Note: The sample includes all papers (9,429). The coefficients measure media matches in the first 4
weeks after the NBER release date (release period) compared to the 4 weeks before (control period).
All regressions control for levels of included characteristics, date of release, number of authors and
JEL categories. Journal impact factors and ranks are set to the minimum if no (ranked) publication
exists. Standard errors reported in parentheses are clustered by paper. I( ) denotes indicator
functions. ∗ significant at 10% level, ∗∗ significant at 5% level, ∗ ∗ ∗ significant at 1% level.

                                                                    25
Table A.4: Media reports by number of working paper versions (Restricted sample)

                                              (1)          (2)         (3)           (4)
         WP released

         × # other WP versions             0.011**       0.006
                                           (0.005)      (0.005)
         × 1 other WP version                                        0.027         0.019
                                                                    (0.018)       (0.018)
         × 2 other WP versions                                      0.045**       0.034*
                                                                    (0.019)       (0.019)
         × 3+ other WP versions                                     0.046**        0.031
                                                                    (0.021)       (0.021)
         JEL Categories                                    X                         X
         Note: The sample includes all multi-authored papers (9,429). The coeffi-
         cients measure media matches in the first 4 weeks after the NBER release date
         (release period) compared to the 4 weeks before (control period). All regres-
         sions control for levels of included characteristics, date of release and number
         of authors. Standard errors reported in parentheses are clustered by paper. ∗
         significant at 10% level, ∗∗ significant at 5% level, ∗ ∗ ∗ significant at 1% level.

                                                 26
Table A.5: Author-level media reports by author characteristics (Restricted sample)

                                                (1)          (2)        (3)           (4)            (5)
WP released

× Gender - Female                             -0.007        -0.009     -0.005       -0.009         -0.006
                                             (0.007)       (0.008)    (0.008)      (0.008)        (0.008)
× Gender - Unknown                            -0.001        -0.004     0.000        -0.001          0.001
                                             (0.012)       (0.012)    (0.011)      (0.012)        (0.011)
× Experience (10−1 ×years)                                  -0.006   -0.023***    -0.011***      -0.025***
                                                           (0.004)    (0.005)      (0.004)        (0.005)
× ln(Author article views+1)                                         0.016***                    0.014***
                                                                      (0.004)                     (0.004)
× I(Author article views observed)                                   -0.140***                   -0.122***
                                                                      (0.039)                     (0.037)
× ln(School citation score)                                                         0.009**        0.008*
                                                                                    (0.005)       (0.004)
× I(School citation score observed)                                                  0.018*         0.017
                                                                                    (0.010)       (0.010)
Note: The sample includes all author-paper combinations of multi-authored papers (26,853). The coeffi-
cients measure media matches in the first 4 weeks after the NBER release date (release period) compared
to the 4 weeks before (control period). All regressions control for levels of included characteristics, date
of release, number of authors and JEL categories. Experience is proxied by years since first NBER pub-
lication. Article views and citation scores are set to the respective minimum if author/school data not
available. I( ) denotes indicator functions. ∗ significant at 10% level, ∗∗ significant at 5% level, ∗ ∗ ∗
significant at 1% level.

                                                      27
You can also read